The model for improving big data sub-image retrieval performance using scalable vocabulary tree based on predictive clustering

摘要

Scalable vocabulary tree (SVT) is a data compression structure which gains scalable visual vocabularies from hierarchical k-means clustering of local image features. Due to both high robustness in data retrieval and great potentials to process huge data, it has become one of the state-of-the-art methods building on the bag-of-features. However, such bag-of-words representations mainly suffer from two limitations. The paper gives a performance research of re-ranking in sub-image retrieval using SVT which is built from local Speed Up Robust Features descriptors. Firstly, the paper gives a study on retrieval performance using different single layers of the tree, which tells it divides data too coarsely for low layers with a small quantity of leaf nodes, while too finely for the 6-th layer with too many leaf nodes. Then using the best selected layer, the authors give a comparative analysis with popular advanced re-ranking strategies in the existing literatures. Finally, the authors propose a weighted score method that calculates matching score from dominating layers. The experimental results prove that the weighted score method achieves almost optimal retrieval performance when using SVT for data representations. Meanwhile, it almost doesn't increase any computational complexity, and can be implemented easily.